Open AccessImage segmentation of skin cancer using MobileNet as an encoder and linknet as a decoder
Author(s) -
M Widiansyah,
S Rasyid,
P Wisnu,
Agustinus Deddy Arief Wibowo
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1943/1/012113
Subject(s) - segmentation , encoder , computer science , skin cancer , block (permutation group theory) , artificial intelligence , image segmentation , stage (stratigraphy) , process (computing) , cancer , pattern recognition (psychology) , medicine , mathematics , paleontology , geometry , biology , operating system
Cancer causes many deaths in the world. One of the most dangerous types of cancer is skin cancer caused by cancer cells that grow and spread uncontrollably in the skin layers. There are an estimated that 100,000 new cases of skin cancer will occur in the United States, which will cause around 6000 deaths by 2020. The survival rate of people with skin cancer can increase if successfully treated at an early stage. Examination of the lesion or dermoscopy images takes a long time and is prone to errors due to differences of opinion by doctors. A computer program for image processing was developed to improve accuracy and efficiency that helps doctors diagnose and evaluate. The most crucial stage in the diagnosis process is the segmentation of skin lesions. Wrong segmentation results will affect accuracy at the classification stage. This study developed a segmentation model for a skin cancer image that uses the MobileNet model as an encoder block and Linknet model for a decoder block by selecting the best hyper-parameter value from several training scenarios. MobileNet encoder and Linknet as a decoder are known to have the same ability, which is to overcome the problem of computational efficiency without compromising accuracy. Four hyper-parameters used are learning rate, the number of epochs, image size, and using pre-trained or not. Besides, there is a performance comparison scenario between the MobileNetV1 and MobileNetV2 encoder blocks. The model trained using the 2017 ISIC Challenge dataset consisting of 2000 training data, 150 validation data, and 600 test data. The Intersection over Union score obtained was 71.5% from models trained with a hyper-parameter learning rate of 0.001, 50 epochs, with an image size of 256 × 256 × 3, using a pre-trained model, and using a MobileNetV1 encoder block. The mask image that is the result of the segmentation of skin cancer images using the MobileNet encoder model and Linknet decoder is accurate enough to be used in the process at a later stage.